Professional Documents
Culture Documents
Introduction
The collection and study of retail transaction data, known as market basket
analysis, has become increasingly prevalent in the past several years. Many
supermarkets, for example, issue loyalty cards [27]. While providing discounts to
the customer, these cards allow the retailer to develop a better understanding of
individuals purchasing habits by associating customers with transactions. The
uses of this information vary, but may include informing product placement
decisions, designing personalized marketing campaigns, and determining the
timing and extent of product promotions [1, 2, 14] among others.
Formally, the task of market basket analysis is to discover actionable knowledge in transaction databases. The problem can be understood as follows: A
standard retail store sells a large set of products P. Define a transaction p P
as the set of products an individual customer buys in a single trip to the store.
The stores transaction database T = {p} is the set of all transactions the
store has processed within a given time period. Ultimately, an effective analysis method should enable the retailer to draw clear, comprehensive conclusions
from the data.
One popular tool for market basket analysis in practice is the mining of
association rules [2]. A set of association rules R(T, s, c) is defined by a transaction database T, a minimum support parameter s and a minimum confidence
parameter c. Define A and B as arbitrary sets of products. Further, define A
(analogously B) as the set of transactions containing every product in A (B).
Formally, R is the set of all rules A B such that:
1. |AB|
|T| s
2. |AB|
|A| c.
Association rules have found successful application in many diverse contexts
and a number of algorithms have been developed to discover them efficiently
[2, 10, 23, 43], but they are not without limitations. The most prominent of
these is sheer volume. Large transaction datasets tend to contain hundreds or
thousands of rules at reasonable levels of support and confidence, and many
of these may be redundant or obvious [25]. As a result, it is often difficult to
isolate interesting relationships.
Two distinct classes of methods have evolved to address this problem. One
class [20, 25, 40, 41] attempts to eliminate any rules that may be redundant,
while the other [18, 28, 35] aims to elevate rules that are especially interesting
(by sorting on an objective measure). Unfortunately, the concepts of both interestingness and redundancy are somewhat subjective. As a result, (which we
show in Section 2) these methods are of limited use in practice.
Ultimately, existing literature on market basket analysis has failed to provide
conclusive answers to some of the fields most pressing questions. For example,
there is no widely-accepted means of isolating representative or useful relationships in market basket datasets and no existing work of which we are aware has
attempted to offer any manner of procedural guidance for analyzing such data.
In other words, no work has addressed the question Given a new market basket dataset, what method or methods should I apply in order to obtain effective
insights?
This work attempts to address these concerns and improve the power and
clarity of market basket analysis by modeling transactional data as a network.
We show that by detecting communities of products in this network, we can
discover strong and expressive relationships among products including relationships that are difficult to discover with traditional association rules. We then
build on our experience with product networks and with a number of different
market-basket and graph-theoretic algorithms to propose a novel procedure for
mining unseen market basket datasets. The network representation of transaction data allows for the use of a diverse array of algorithms previously unavailable to the association rule community. As a result, this procedure is the first
comprehensive market basket analysis framework ever proposed in the literature. All of our developments and conclusions are verified on real transaction
data, consisting of over 660,000 transactions across more than 2,200 items, from
an on-campus convenience store at the University of Notre Dame.
The remainder of the paper is organized as follows: Section 2 explores the
strengths and weaknesses of traditional association rules analysis on our transaction data. The results presented here motivate the rest of the paper and serve
as an introduction to the data itself. Section 3 introduces the concept of product networks and presents some properties of our network. Section 4 describes
our community detection approach to market basket analysis and presents the
first known interestingness measure for communities of products. Section 5
develops a comprehensive and novel framework for market basket analysis, incorporating both techniques introduced in this paper and previously-developed
network analysis methods. Finally, Section 6 acknowledges some related work
not mentioned elsewhere in the paper and Section 7 concludes.
Association Rules
A popular approach for analyzing market basket data is the discovery and interpretation of association rules. The association rules problem [2] is defined as
follows:
Given a threshold s, called the minimum support and a threshold c, the
minimum confidence, find all rules of the form A B, where A and B are sets
of products, such that:
1. A and B appear together in at least s% of transactions.
2. B occurs in at least c% of the transactions in which A occurs.
Sets of products are typically called itemsets, itemsets of size k are called
k-itemsets, and sets that meet the minimum support criterion are typically
called large or frequent itemsets. An association rule is said to be supported
in a transaction database if it meets both the minimum support and minimum
confidence criteria.
Algorithms for efficiently enumerating association rules are well-known
[2, 21, 42] and are a popular tool for unsupervised data exploration. As they
came into widespread use, researchers noticed that understanding the rules
themselves was not a trivial matter. First, there is no obvious method for
choosing appropriate support and confidence thresholds. If the thresholds are
chosen too high, interesting associations may be missed. However, if they are
chosen too low, the user may be inundated with thousands of weak rules that
P (AB)
P (A)P (B)
(1)
Table 1: High, Low, and Mean rank and Standard Deviation of Ranks for the
top 10 rules by average rank among the 21 interestingness measures in [35]
Rule
CREAM CHEESE BAGEL
Cake Mixa Frosting
VAULT SODA VAULT ZERO
YORK MINT PATTIES, DIET COKE 20 OZ NEWSPAPER CHICAGO TR
NEWSPAPER CHICAGO TR, DIET COKE 20 OZ YORK MINT PATTIES
BAGEL CREAM CHEESE
CREAM CHEESE, COFFEE 12 OZ BAGEL
NYQUIL DAYQUIL
VAULT ZERO VAULT SODA
Frosting Cake Mix
a Product
High
1
3
6
2
8
1
1
1
16
3
Low
128
65.5
71
96
96
129
133
118.5
70
69
Mean
18.07
21.85
24.95
28.05
28.85
31.37
32.02
33.35
33.40
34.37
StDev
33.79
19.61
15.77
25.67
22.90
36.32
42.90
33.92
13.41
43.42
size in market basket data. For our data, there are no hyperclique patterns of
size greater than two, even at support as low as 0.005%. Therefore hyperclique
patterns, while effective at discovering certain strong relationships, are hardly
a sufficient analysis technique on their own.
Association Rules Networks [12, 13, 33] reduce the ruleset by focusing solely
on rules related to a single product. More specifically, given a set of association
rules R and a target product z, the association rules network ARN(R, z) is the
unique directed hypergraph G satisfying the following properties:
1. Any hyperedge in G corresponds to a rule in R with a one-item consequent.
2. There is a hyperedge corresponding to a rule whose consequent is the
target product z.
3. The target product z is reachable from every vertex v in G.
4. No vertex v 6= z is reachable from z.
Generally speaking, an ARN shows the extent to which rules flow into the
target product. The resulting network can show both direct and indirect associations of the target product z. However, Association Rules Networks can be quite
sensitive to the choice of target product, and there is no obvious proper choice.
As a result, one must have some idea of the products he or she is interested in
before association rules networks are applicable. We explore the integration of
association rules networks into a broader strategy for market basket analysis in
Section 5.3.
The above discussion suggests that no technique currently available in the literature sufficiently addresses the problem of finding meaningful relationships in
large transaction databases. This deficiency motivates our discussion of network
methods for market basket analysis, which is the subject of the next section.
We do not claim to definitively solve the market basket problem. However, we
will show that as a first exploratory step, our techniques can discover expressive
relationships from which we can draw direct conclusions about the nature of
customer behavior in a store.
(a)
(b)
Figure 2: Degree distribution for (a) the entire network and (b) the neighbors
of a single product.
products. Networks based on citations or phone calls, for example, do not suffer
this problem to nearly the same degree.
In citation networks, two nodes linked together by an edge are necessarily
related: if one paper cites another, there is a reason. A cell phone network
will have a small number of incidental links, (wrong numbers, telemarketing,
or random personal business), but most of the time, when one person calls
another, it implies a connection between them. Product networks are different.
Simply because a person buys paper towels and spaghetti sauce in the same
transaction does not entail a common motivation for the two purchases. Worse,
a person who buys several unrelated items in a single transaction will form a
clique among them, despite the absence of any true relationship.
As a result, product networks are very dense, with a large number of connections per node, but many of these edges are meaningless: representing spurious
associations generated by chance. Our network contains 2,248 products and almost 250,000 edges between them. However, over 150,000 of these edges have a
weight of one, meaning the two products were bought together only once in the
entire year 2006, and over 235,000 have weight less than 10. These extremely
low-weight edges are common and are unlikely to represent strong relationships.
One natural consequence of this density, many popular network statistics are
unusually skewed. For example our product network has a 90% effective diameter of 4 and a full diameter of 5, much smaller than we would expect in a social
network of the same size, and the average clustering coefficient is relatively high
at 0.518.
In order to remove some of the noisy edges created by coincidental purchases
and improve the quality of our subsequent analysis, we establish a minimum
threshold , such that an edge exists between two products only if they have
been bought together at least times. This is analogous to choosing a minimum
support threshold for association rules. Note that, in the pruned network, the
weight of the any remaining edge is unchanged.
Having described the construction of a product network and studied some of
its properties, we now turn our attention to the analysis of the product space.
Since the primary focus of market basket analysis is the discovery of relationships between products, we need to find groups of products whose structure or
position within the network reveals useful information about the store itself.
Many real-world interaction networks naturally contain communities:
groups of nodes that are more strongly connected to each other than they are to
the rest of the network. Often, these communities have an easily-interpretable
significance. In a cell phone network [34], for example, communities may represent families or circles of friends. Conversely, in a network of web pages [24]
they may represent sites devoted to a common interest or theme. Community
detection has been applied successfully in a numerous fields of science, ranging
from social network analysis [34] to biology [3] and molecular physics [26]. It
seems logical to expect that communities of products, since they are mutually
strongly-connected, would be of particular interest. Therefore, the remainder
of the paper will focus on the problem of community detection in product networks, and show how communities of products can be used to gain insight in to
9
where eii is the fraction of edges that join vertices in community i to other
vertices in community i and ai is the fraction of edge endpoints that lie in
community i. Modularity measures the difference between the number of incommunity edges in a given set of communities and the expected number of
in-community edges in a random network with the same degree distribution.
This notion is very intuitive. If a set of communities has a large fraction of
its edges falling within communities, (and therefore a relatively small fraction
falling between communities), then that particular community decomposition
probably represents a strong community structure.
The application to market basket analysis is clear: isolating tightlyconnected communities within the network of products will allow us to identify
strong relationships among the products and, therefore meaningful correlations
in customer purchase behavior. Furthermore, because communities can be arbitrarily large, they should be able to represent these relationships much more
expressively and with less redundancy than ordinary association rules.
4.1
Before we present our results, we quantify the utility of a community. Specifically, we wish to answer the question: given a set of communities in a product
network, which are most useful to a human analyst?
Intuitively, the utility of a community can be determined by two opposing
forces: information, and information density. A useful community will be large
enough to provide a substantial insight into customer behavior, but small enough
to be human-interpretable. To this end, we propose the following quantitative
definitions. Define the information present in a community to be the sum, over
all the edges in the community, of the confidence of the relationship indicated by
the edge. The confidence of the relationship A B is the observed conditional
probability that B is purchased given that A is purchased.
X
I(Gi ) =
P (p1 |p2 )
(3)
(p1 ,p2 )Ei
10
2I(Gi )D(Gi )
.
I(Gi ) + D(Gi )
(5)
Substituting the definitions of I(Gi ) and D(Gi ) into Equation 5 yields: U (Gi ) =
i|
D(Gi ) |V|V
. Thus, our measure prefers dense communities but given two comi |+1
munities of roughly equal density, it favors the larger one. This matches the
intuition given earlier.
Because the computation in Equation 3 depends on the actual number of
edges present in the community, our utility measure depends somewhat on the
method of graph construction. In other words, if we allow an edge between any
two products that are bought together, the computation will be different than if
we restrict edges to products bought together at least 100 times. The end result
of this is that our utility measure is not comparable across different network
constructions. We do not consider this to be a significant issue because it is
designed to help a human analyst assess one set of communities.
While our utility measure is designed for product networks, we believe that
the tradeoff between size and density is very general and that, in principle,
Equation 5 could be applied to other domains. In an email network, for example,
if one defines information as the frequency of email correspondence between
members of the community over some time period, an analog of Equation 5
follows naturally.
4.2
Figure 3: The first two communities in our data, ranked by the measure given
in Equation 5.
chips (and respectively salsa) are substitutes for one another. The salsa con
queso is an exception, because it is distinct from the other types available.
Figure 3(b) shows the second-ranked community, a collection of eggs and
baking products. The structure of the community, with eggs (EGGS CSPRING
8CT) as a hub in the center and the baking items the periphery, seems to imply
that when people buy eggs in our store, they buy them for baking. Further
investigation supports this initial hypothesis.
There were 541 distinct products bought with EGGS CSPRING 8CT at our
store in the calendar year 2006, and in 18.5% of the cases, they were bought
alone. However, at least one item among the six neighbors appears in over 39%
of all transactions containing EGGS CSPRING 8CT, which is especially significant
because most of the transactions in our store are small. As a case study, we
further quantify the impact of this particular community. Similar analysis can
be applied to other communities, but space limitations preclude such analysis
in this paper. Intuitively, cake mix is the most likely causal item in the group
(it is unlikely, for example, that people buy frosting because they have a craving
for eggs). Therefore, we calculate expected additional sales from each sale of
cake mix as:
E(Sales) =P (Eggs|CakeM ix) P rice(Eggs)
+P (F rosting|CakeM ix) P rice(F rosting)
and find that the store can expect to generate $2.30 in additional sales from
each cake mix sold. Therefore, the store stands to profit from any promotion
that increases the sales of cake mix at a cost of less than $2.30 per transaction.
Since cake mix itself costs $2.69, the expected additional revenue is 85.5% of the
items purchase price. This analysis is admittedly simple, but it demonstrates
that communities can help identify profitable promotions in a store.
The third-and-fourth-ranked communities, shown in Figures 5(a) and 5(b)
are communities of cereal and milk. The first of these shows a small container
of milk as a hub surrounded by a series of cereals. In this case, the milk is small,
at one pint, and many of the cereals are smaller individual-serving cereals. The
second is composed of two nearly-disconnected subgraphs: a hub-and-spoke
arrangement of larger milks and cereals and a clique of sodas. The disparate
structures are each connected, by one edge, to a single product: plastic cups.
These communities support several conclusions in addition to the notion
that people buy cereal and milk together. First, there are separate relationships
between cereal and milk at two levels: smaller sizes of milk correlate with smaller
sizes of cereal, while larger milks relate to larger cereals. Second, the strong
mutual correlation among sodas suggests that they are often purchased several
at a time, while the disconnection among cereals indicates that people buy them
largely for personal use.
The final community of interest is shown in Figure 5(c): a community containing fruit, salad, yogurt. It is much less dense than the others and therefore,
at number eight, is ranked much less favorably. However, it still contains useful
insights. Figure 5(c) shows the single fruit product (diamond) connected to nine
13
14
among important but more peripheral products. Then, the subsequent association rules analysis can focus more intently on products whose role is not clear
within the community decomposition. The next section describes in greater
detail our proposed framework for such an analysis.
A great deal of literature has been published on the subject of market basket
analysis and survey papers about algorithms [23, 43], interestingness measures
[28, 35], and visualization techniques ([5], section 2) abound. In spite of all
this effort, however, the community has made no substantive attempt to answer
the following basic question: Given a fresh, unseen market basket dataset what
method or set of methods should be employed to obtain quick, actionable results?
There are several possible reasons for this. The first is a dearth of widelyavailable transaction data, which we alluded to in the introduction. The second
is a general lack of diversity in analysis techniques: maximal itemset mining,
for example, is not different enough from traditional association rules such that
the techniques can be complementary, with one strong where the other is weak.
Finally, most studies that do consider real data are only conducted within a
single domain (i.e. supermarkets or online retailers), and so the ability to draw
overarching conclusions is limited.
Since we too are confined to a single dataset, we cannot address the third
concern, but this section addresses the first and the second. In doing so, we call
upon not only the techniques developed here in Section 3, but also a series of
methods developed by other authors. To our knowledge, these methods (Association Rules Networks [12, 13, 33] and Center-Piece Subgraphs [36]) have not
been generally applied to market basket data, but in the course of our work we
have found that they complement community detection nicely.
The rest of the section is organized as follows: Section 5.1 explores practical
concerns regarding the use of Association Rules Networks (introduced in Section 3), Section 5.2 introduces the Center-Piece Subgraph problem and studies
its application in the domain of product networks, Section 5.3 ties together the
discussion of this section and the prior one in order to propose a unified strategy
for mining market basket data, and Section 5.4 briefly discusses strategies for
parameter selection.
5.1
Recall from Section 3 that an Association Rules Network ARN (R, z) is a directed hypergraph representation of the ruleset R that mops out the direct and
indirect associations of the target product z. The concerns we must address
when applying Association Rules Networks are 1) How do we choose an appropriate ruleset R? and 2) How do we choose an appropriate item z? The
first question essentially boils down to the appropriate choice of support and
confidence parameters, and we do not address it here. With regard to the sec-
15
(a) Association Rules Network with z = eggs (b) Association Rules Network with z = cake
(EGGS CSPRING 8CT)
mix (DH YELLOW CAKE MX 18)
16
17
ation Rules Network from transaction data, is to choose the item that appears
in the most rules in the underlying ruleset R. One might consider instead the
most popular product in the store, or the item which has been bought with the
greatest number of other products. In our data, however, these strategies are
less effective. BULK CANDY, which is both the most frequently-sold and bought
with the most items, has only two products in its Association Rules Network,
and one popular type of water (WATER DASANI 20 OZ), has none.
The reason for this is that association rules involving BULK CANDY and WATER
DASANI 20 OZ, which are bought with a stunningly wide variety of items, do
not meet the minimum confidence criterion that we have used throughout the
paper. We contend, however, that relationships which do not meet the minimum
confidence criterion may still be interesting. There are several potential causes
of low confidence, but the most relevant in the case of water is substitution.
There are many different types of water available in the store, and this variety
erodes the confidence of certain relationships.
To illustrate the effect of substitution on rule confidence, assume n different
products F1 , . . . , Fn are all substitutes for each other, meaning that they serve
roughly the same function F . Furthermore, assume a product P correlates with
items of the function F , such that the confidence of the association rule F P
is c or
|F P|
= c.
|P|
(6)
If the products F1 , . . . , Fn are all bought equally with P, then for any Fi ,
the confidence of the rule Fi P is given by
|F P|
n
|P|
c
.
n
(7)
5.2
Center-Piece Subgraphs
18
Define a Random Walk with Restart (RWR) [37] on the graph G starting from
a node n V (G) as follows: At time t, a randomly-walking particle existing
at node nt V (G) (n0 = n) transmits itself to one of the neighbors of nt with
a probability proportional to the weight of its edge with nt . At any time, the
particle has a fixed probability c of returning to node n.
From the normalized matrix of edge weights W, one can calculate the prob(t)
ability pi,j that a randomly-walking particle starting at node i stands at j after
(t)
(8)
(10)
vV (H)
Ref. [36] provides a fast algorithm for extracting subgraphs with high g(H),
and our experience shows that it scales to networks with thousands of nodes.
In the next section we explore practical concerns regarding the application of
CePS to market basket analysis and present results from our data.
5.2.2
Each technique we have discussed to this point has been limited by the need to
specify a minimum support (and possibly minimum confidence) with which to
discover relationships. As a result, strong relationships with low levels of support
and substitution relationships with artificially low confidence are undiscovered.
20
Because center-piece subgraphs are constrained in size by the budget parameter b, it is unnecessary to further constrain them with minimum support
and confidence parameters. As a result, they are the only technique we have
discussed which is capable of discovering relationships between any and all products that make up the product space. The remainder of the section will show
that this property makes center-piece subgraphs invaluable for the exploration of
results obtained through other means. Specifically, they are effective for either
verifying hypotheses suggested by other techniques or explaining relationships
that do not, on the surface make sense.
Figure 8(a) shows a center-piece subgraph constructed from the full 2006
product network using a type of tortilla chips (TOSTITOS SUPER SIZE) as the
query node and a budget b = 10. The network contains other chips and salsa,
as our prior experience would lead us to expect, but also contains some items
(BULK CANDY and BAGEL) that are marginally related at best. We explored this
phenomenon by constructing subgraphs of gradually increasing size in order to
determine which items the algorithm considered more important with respect
to the tortilla chips. In doing so, we found that the BULK CANDY was added as
the 6th member of the subgraph, before other products to which the chips have
a stronger connection.
The reason for this is that BULK CANDY, as a popular product, is bought with
a tremendously large array of other products (recall the degree distribution of
Section 3). To see why this causes problems for CePS, imagine a seldom-sold
product pj , appearing in 5 transactions, with which BULK CANDY is bought once.
By standard normalization, the transition probability from pj to BULK CANDY is
at least 1/5, meaning that any random particle that reaches j is highly likely to
reach BULK CANDY. Combining this effect over hundreds of less popular products
results in a very substantial steady-state probability for popular products.
To reduce the influence of such products, we weighted the edges by confidence
instead of by absolute support. That is, the edge A B is weighted with
min(P (A|B), P (B|A)). There are two distinct advantages to using confidence
in this instance. First, it forces all edge weights onto a uniform scale between
zero and one. Second, it lessens the impact of coincidental purchases with
popular products. In the example of the previous paragraph, the weight of the
1
edge between pj and BULK CANDY is now 60,000
and after normalization it is
likely that the transition probability from pj to BULK CANDY is much lower.
Figure 8(b) shows the impact of weighting edges by confidence. Now, instead
of extraneous products like BAGEL and BULK CANDY, we see sodas and other types
of chips, which much more closely matches our intuition and corroborates the
results found with other techniques.
Figure 9 shows a center-piece subgraph with eggs (EGGS CSPRING 8CT) as
the lone query node and a budget of 10. When we examined the community of
eggs and cake mix in Section 3 we concluded that when customers bought eggs
in our store, they bought them for baking. The subgraph in Figure 9 further
corroborates this notion: it includes four additional products (brownie mix,
butter, margarine, and chocolate chips) and all of them are baking products.
To this point, we have used CePS simply to explore the neighborhood of
21
Figure 9: A center-piece subgraph with eggs (EGGS CSPRING 8CT) as the query
node.
individual items, similar to the way in which we might apply Association Rules
Networks. As we mentioned before, however, the CePS algorithm is actually
much more general, and can handle any number of query nodes. The following
discussion explores the ability of CePS to explain a single association rule.
Figure 10 shows a ten-node center-piece subgraph for one of the less intuitive (and more interesting) rules in the dataset: DIET COKE 20 OZ, YORK MINT
PATTIES NEWSPAPER CHICAGO TR. Specifically, it is a center-piece subgraph
with those three items as query nodes and a budget of 10. The three items
in question seem to be entirely unrelated, and yet the rule is ranked highly by
a number of interestingness measures (Table 1). Ideally, the center-piece subgraph would illuminate the relationship between the products and explain the
association.
Looking at the network, we see something interesting. In addition to patties
and Kit Kat, which appear in the Association Rules Network of Figure 7, we
also see three more types of candy: Hersheys, Mounds and Chuckles. This
observation implies that there is some sort of relationship between Chicago
Tribune, Diet Coke, and candy. As it turns out, the newspapers in our store are
located at the front of the store, next to the rack where those candies are sold.
Figure 10 shows that, because center-piece subgraphs can consider the entire
product network without requiring excessive computation time or providing
overwhelming output, they are very effective for exploration or validation of
relationships provided by other methods. As such, they complement nicely the
other techniques outlined in this paper.
Center-Piece Subgraphs require substantially more parameters than any of
the other techniques we have discussed. All of our experiments were conducted
on small networks (b 10), with hard AND, meaning that k is equal to the
number of query nodes. Though we did not conduct any detailed studies of
22
Figure 10: A center-piece subgraph with Diet Coke (DIET COKE 20 OZ), Newspaper (NEWSPAPER CHICAGO TR), and Peppermint Patties (YORK MINT PATTIES)
as query nodes, to explain the association rule.
the parameter selection process, informally we found that the choice of k and b
makes little difference in the quality of the subgraph discovered. By choosing b
to be large, we observed that popular products such as BULK CANDY and BAGEL
came to be included in the subgraph. Altering k had no discernible effect for
the types of queries we tried.
5.3
The research we present here has allowed us to make and corroborate a number
of significant observations about market basket analysis of real-world data. We
re-state the chief observations here, citing the work of others where appropriate.
1. Deriving interesting, actionable knowledge from association rules is difficult because rulesets are often muddied by a preponderance of obvious or
redundant rules [25].
2. One can choose to mine maximal or closed itemsets instead, but these
techniques fail to prune away many redundant rules.
3. Similarly, one may choose to rank rules by an interestingness measure,
but there are many such measures to choose from and they may rank rules
inconsistently [35]. As such, it may be difficult to choose an appropriate
measure in the absence of prior knowledge.
4. Detecting communities of products within the network formed by customer purchases can alleviate redundancy by discovering larger, more expressive relationships among groups of products. However, community
detection is less effective within the dense core of the network and requires
a minimum support threshold, which imparts parameter sensitivity.
23
5.4
Since the first step in our proposed procedure requires the user to choose a
minimum support parameter, we attempt to provide some guidance into this
choice. We are aware of no prior work from which to draw, but one can imagine
several reasonable options. For example, one might select an arbitrarily high
threshold and iteratively reduce it until the number of rules becomes unmanageable. Alternatively, one may attempt to find a certain number (some hundreds
or thousands) of rules, or a certain number of rules that score highly based on
his or her favorite interestingness measure.
All of these are valid choices and to evaluate them critically is beyond the
scope of this work. However, if community detection is the target then existing
community detection research affords us another option. In Section 3, we briefly
alluded to the fact that community detection algorithms find poor communities
at low levels of minimum support. This fact can be used, in principle, to choose
24
Related Work
Conclusion
This work deals primarily with the application of network techniques to the
problem of market basket analysis: the location of meaningful associations in
customer purchase data. There is an overwhelming abundance of prior research
in the mining of mining market basket data in general, and the use of association rules in particular. The bulk of this research has focused on developing
algorithms for mining association rules [2, 10, 9, 42, 43], techniques for visualizing association rules [5, 22, 25, 38], techniques for eliminating redundant rules
[25, 20, 40, 41], objective measures of association interestingness [18, 28, 35],
or comparing the performance of association rule algorithms on either real or
synthetic datasets [23, 44]. However, there has not been much work from a
practitioners view point towards answering: Given an unseen market basket
dataset, what set of steps should I follow to conduct a thorough, complete analysis? Our work provides a comprehensive framework aimed at answering this
question.
First, we study the properties of networks of products and show that detecting communities within these networks can uncover expressive relationships
between products that may be difficult to find with association rules. We show
that, in addition to being more expressive than association rules (in that rela-
26
Acknowledgments
This work partially supported by the National Science Foundation under grant
NSF 0826958, the NET Institute, and the Arthur J. Schmitt Foundation.
References
[1] G. Adomavicius and A. Tuzhilin. User profiling in personalization applications through rule discovery and validation. In Proceedings of KDD, pages
377381. ACM New York, NY, USA, 1999.
[2] R. Agrawal and R. Srikant. Fast algorithms for mining association rules in
very large databases. In Proceedings of the 20th International Conference
on VLDB, pages 487499, Santiago, Chile, 1994.
[3] S. Asur, D. Ucar, and S. Parthasarathy. An ensemble framework for clustering protein-protein interaction networks. In ISMB/ECCB, pages 2940,
2007.
[4] A. Barabasi and E. Bonabeau. Scale-free networks. Scientific American,
288(5):509, 2003.
[5] J. Blanchard, F. Guillet, and H. Briand. Exploratory visualization for
association rule rummaging. In KDD-03 Workshop on Multimedia Data
Mining (MDM-03), 2003.
[6] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre. Fast unfolding of communities in large networks, 2008.
27
[7] T. Brijs, K. Vanhoof, and G. Wets. Defining interestingness for association rules. International journal of information theories and applications,
10(4):370376, 2003.
[8] S. Brin, R. Motwani, L. Page, and T. Winograd. What can you do with a
Web in your Pocket? Data Engineering Bulletin, 21(2):3747, 1998.
[9] S. Brin, R. Motwani, and C. Silverstein. Beyond market baskets: generalizing association rules to correlations. Proceedings of the ACM SIGMOD,
pages 265276, 1997.
[10] S. Brin, R. Motwani, J. Ullman, and S. Tsur. Dynamic itemset counting and implication rules for market basket data. ACM SIGMOD Record,
26(2):255264, 1997.
[11] L. Cavique. A scalable algorithm for the market basket analysis. Journal
of Retailing and Consumer Services, 14(6):400407, 2007.
[12] S. Chawla, B. Arunasalam, and J. Davis. Mining open source software (oss)
data using association rules network. PAKDD, pages 461466, 2003.
[13] S. Chawla, J. Davis, and G. Pandey. On Local Pruning of Association
Rules Using Directed Hypergraphs. 20th International Conference on Data
Engeneering, 2004.
[14] Y. Cho, J. Kim, and S. Kim. A personalized recommender system based
on web usage mining and decision tree induction. Expert Systems with
Applications, 23(3):329342, 2002.
[15] A. Clauset, M. Newman, and C. Moore. Finding community structure in
very large networks. Phys. Rev. E, 70(066111), 2004.
[16] A. Clauset, C. Shalizi, and M. Newman. Power-law distributions in empirical data. axriv, 706, 2007.
[17] N. Du, B. Wu, X. Pei, B. Wang, and L. Xu. Community detection in largescale social networks. In Proceedings of WebKDD, pages 1625. ACM,
2007.
[18] W. DuMouchel and D. Pregibon. Empirical bayes screening for multi-item
associations. Proceedings of KDD, pages 6776, 2001.
[19] B. Fonseca, P. Golgher, B. Possas, B. Ribeiro-Neto, and N. Ziviani.
Concept-based interactive query expansion. In Proceedings of CIKM, page
703. ACM, 2005.
[20] K. Gouda and M. Zaki. Efficiently mining maximal frequent itemsets. In
Proceedings of ICDM, pages 163170. IEEE Computer Society, 2001.
28
[21] J. Han and J. Pei. Mining frequent patterns by pattern-growth: methodology and implications. ACM SIGKDD Explorations Newsletter, 2(2):1420,
2000.
[22] M. Hao, U. Dayal, M. Hsu, T. Sprenger, and M. Gross. Visualization
of directed associations in e-commerce transaction data. Proceedings of
VisSym, 1:185192, 2001.
[23] J. Hipp, U. G
untzer, and G. Nakhaeizadeh. Algorithms for association
rule mininga general survey and comparison. ACM SIGKDD Explorations
Newsletter, 2(1):5864, 2000.
[24] J. Kleinberg and S. Lawrence. The structure of the web. Science, 294:1849
1850, 11 2001.
[25] M. Klemettinen, H. Mannila, P. Ronkainen, H. Toivonen, and A. Verkamo.
Finding interesting rules from large sets of discovered association rules.
Proceedings of CIKM, pages 401407, 1994.
[26] C. Massen and J. Doye. Identifying communities within energy landscapes.
Physical Review E, 71(4):46101, 2005.
[27] C. Mauri. Card loyalty. A new emerging issue in grocery retailing. Journal
of Retailing and Consumer Services, 10(1):1325, 2003.
[28] K. McGarry. A survey of interestingness measures for knowledge discovery.
The knowledge engineering review, 20(01):3961, 2005.
[29] M. Newman. Detecting community structure in networks. The European
Physical Journal B-Condensed Matter and Complex Systems, 38(2):321
330, 2004.
[30] M. Newman. Finding community structure in networks using the eigenvectors of matrices. Physical Review E, 74(3):36104, 2006.
[31] M. Newman and M. Girvan. Finding and evaluating community structure
in networks. Physical Review E, 69(2):26113, 2004.
[32] C. Palmer and C. Faloutsos. Electricity based external similarity of categorical attributes. Lecture notes in computer science, pages 486500, 2003.
[33] G. Pandey, S. Chawla, S. Poon, B. Arunasalam, and J. Davis. Association
Rules Network: Definition and Applications. Statistical Analysis and Data
Mining, 1(4), 2009.
[34] K. Steinhaeuser and N. Chawla. Community detection in a large-scale real
world social network. In LNCS. Springer Verlag, 2008.
[35] P. Tan, V. Kumar, and J. Srivastava. Selecting the right objective measure
for association analysis. Information Systems, 29(4):293313, 2004.
29
30